Upload
others
View
11
Download
0
Embed Size (px)
Citation preview
CENTER FOR DIGITAL SAFETY & SECURITY
ARTIFICIAL INTELLIGENCE Applications, Services and Solutions
systems to analyse and interpret complex, unstructured data
deriving from images, videos, audio and text files, as well as
sensor data. Perhaps the most important function of AI is,
however, interaction with the physical environment, such as
with robots or sensors for the efficient control and navigation of
highly automated systems, with both human interaction (e.g. via
gestures, speech, facial expressions) and interaction between
machines within M2M constellations.
The following pages provide an overview of state-of-the-art
AI-based services and solutions developed at AIT Center for
Digital Safety & Security together with and for partners in
business and industry.
Headquarter of AIT Austrian Institute of Technology in Vienna, Austria. INTRODUCTIONToday Artificial Intelligence (AI) is applied in almost all
industrial sectors and social subsystems.
It is used in areas including traffic and transport (e.g. autono-
mous driving), robotics, industrial automation (smart factories),
public safety, crisis and disaster management, cybersecurity
(e.g. protection of industrial control systems), in a range of
banking, insurance and legal business applications, as well as
in fields including education, social services, energy, the
environment and healthcare provision. Its primary intention is,
of course, to enhance customer convenience.
The Center for Digital Safety & Security at AIT has comprehen-
sive expertise in data science and AI, and is working to develop
state-of-the-art AI-based services and solutions. The strategic
key areas for new AI technologies lie in dedicated application
areas such as monitoring (e.g. cybersecurity), allowing large
data volumes to be analysed in real time and identifying
patterns as well as deviations. By using self-learning models,
AI systems are also suitable for finding and extracting relevant
information from large volumes of data (data mining), and
interpreting the abstract patterns that have been identified.
Another key function of AI lies in the ability to make predictions,
such as forecasting future trends. The interpretation capability
of AI systems is the most significant factor. This allows AI
3
5
ARTIFICIAL INTELLIGENCE FOR NETWORK ANALYSIS P. 6
NETWORK SECURITY (AI4NETSEC) P. 6
NETWORK ANOMALY DETECTION & DIAGNOSIS (AI4NETADD) P. 6
NETWORK MONITORING AND ANALYSIS (AI4NETMON) P. 7
INTERNET QUALITY OF EXPERIENCE (AI4NETQOE) P. 7
AI FOR NATURAL LANGUAGE PROCESSING (NLP) P. 8
AIT´S ANNOTATION PLATFORM P. 9
MULTI-MODAL AI FOR SECURITY APPLICATIONS P. 10
MULTI-MODAL AI FOR DOCUMENT MANAGEMENT P. 12
MULTI-MODAL AI FOR CUSTOMER SERVICE P. 13
AI FOR PREDICTIVE MAINTENANCE IN INDUSTRY 4.0 P. 14
AI FOR SAFETY & SECURITY IN CYBER-PHYSICAL SYSTEMS P. 15
APPLICATIONS, SERVICES AND SOLUTIONS
ARTIFICIAL INTELLIGENCE FOR NETWORK ANALYSIS Telecommunication networks are the biggest data carriers in
society today, with trillions of events occurring every second,
even in medium-size networks. With such a wealth of big data,
learning-based and AI-enabled networking is a cornerstone of
better, more reliable, and safer networks and services.
However, AI for network data is complex as it involves all of the
major learning and big data challenges (the 4 Vs): massive
volumes of complex and heterogeneous data (Volume and
Variety), fast and highly dynamic streams of data (Velocity), lack
of ground truth for learning (Veracity), as well as highly
unbalanced data, a lack of visibility due to massive adoption of
end-to-end encryption, and more.
AIT can help your organisation unleash the power of AI in a
variety of networking-related domains:
NETWORK SECURITY
(AI4NETSEC)
AI can largely increase the detection accuracy of attacks and
other threats flowing through the network, without sacrificing
the robustness of the analysis (i.e., false alarms). AI can rapidly
identify complex threats and detect previously unseen attacks,
as well as deal with obfuscated and adversarial learning
environments, and consequently helps to drastically improve
your business security. Furthermore, AI-based network
security does not necessarily mean black-box behaviour: you
can still fully track and understand the decisions taken by your
security system, thanks to explainable AI (XAI) concepts and
techniques.
NETWORK ANOMALY DETECTION & DIAGNOSIS
(AI4NETADD)
The automatic detection and diagnosis of the ever-growing
number of anomalies faced by network operators is a para-
mount challenge. By enabling the analysis of highly-dimensio-
nal data with both real-time and large-scale requirements, AI
and big data principles and platforms can significantly improve
the visibility and understanding of such complex and rare
events, empowering a quick troubleshooting process.
6
NETWORK MONITORING AND ANALYSIS
(AI4NETMON)
Network Traffic Monitoring and Analysis (NTMA) is a key
component in network management and is needed to guaran-
tee the correct operation of large-scale and/or complex
networks. Applications such as performance monitoring, traffic
classification and network policing require real-time and
scalable monitoring approaches. Anomaly detection and
security mechanisms require operators to quickly identify and
react to unpredictable events while processing millions of
heterogeneous events. In addition, NTMA systems must be
capable of collecting, storing, and processing massive sets of
historical data for learning and post-mortem analysis. At AIT
we design scalable online and offline data mining and machine
learning-based techniques and platforms to monitor and
characterize extremely large volumes of network traffic data.
INTERNET QUALITY OF EXPERIENCE
(AI4NETQOE)
Quality of Experience (QoE) is a well-known concept that
permits operators to understand and assess the functioning of
networks and services from the standpoint of the end user or
service customer. QoE development has been traditionally
limited to small-scale controlled environments, but today
QoE-based network measurements represent a paramount
source of information for Internet service and content provi-
ders. Here at AIT, we are experts in QoE for networks at scale,
incorporating the QoE paradigm into the design, analysis and
operation of real-world networks, services, and distributed
systems. AI-based Internet-QoE relies on big data analytics to
generate useful user-centric insights from large-scale network
measurements, even under the increasing prevalence of
end-to-end encryption.
7
AIT AUSTRIAN INSTITUTE OF TECHNOLOGY
Dr. Dr. Florian SkopikTel. +43(0) 50550 3116Giefinggasse 4, 1210 Vienna [email protected] www.ait.ac.at/dss
AIT AUSTRIAN INSTITUTE OF TECHNOLOGY
Dr. Pedro Casas
Tel. +43(0) 50550 4104Giefinggasse 4, 1210 Vienna [email protected] www.ait.ac.at/dss
References (excerpt)
ÆCID - Automatic Event Correlation for Incident Detection; https://aecid.ait.
ac.at/
BIG-DAMA – Big Data Analytics for Network Traffic Monitoring and Analysis;
https://bigdama.ait.ac.at/
MobiQoE – Monitoring and Analysis of Quality of Experience in Mobile Net-
works; http://mobiqoe.ait.ac.at/
8
AI FOR NATURAL LANGUAGE PROCESSING (NLP)Discovering valuable information within unstructured, hetero-
geneous text is a challenging task, given the vast amount of
textual data that we must cope with in today’s information
society. AIT provides the expertise and support for a broad
range of use cases by applying a variety of machine-learning
techniques – including algorithmics, mathematics and stati-
stics – to discover topics and semantic relationships, and to
extract meaningful, structured information.
With the advent of more powerful computing hardware, deep
learning has emerged as a new and promising machine-lear-
ning paradigm in natural language processing (NLP). The use
of neural network-based machine learning is not new, but with
the application of new approaches for learning vector repre-
sentations of words in very large text collections, deep learning
has been proven to deliver excellent results in machine-lear-
ning tasks such as classification or the detection of semantic
relationships between words.
While deep learning has changed the manner in which NLP is
carried out, it is only part of the solution. AIT offers advice on
the many classic tasks that still play a fundamental role, such
as corpus building and cleaning, pre-processing, tokenization
and stemming. Classic information retrieval measures such as
co-occurrence statistics, TF-IDF (term frequency–inverse
document frequency), bag-of-words (BoW) representation, or
part-of-speech tagging (PoS) are still important parts of the
NLP toolset. Depending on the specific use case, these are
individually applied as input for Deep Neural Networks (DNN),
or in conjunction with other machine-learning algorithms such
as Support Vector Machines (SVM).
AIT solutions enable the analysis and overview of new text
document collections using unsupervised machine learning
methods. Topic modelling allows us to reveal important
concepts, terms and relationships that occur in a collection of
documents.
9
AIT AUSTRIAN INSTITUTE OF TECHNOLOGY
DI Alexander SchindlerTel. +43(0) 50550 2902Giefinggasse 4, 1210 Vienna [email protected] www.ait.ac.at/dss
References (excerpt)
COPKIT (H2020) – Intelligence-led Early Warning and Early Action system for
Law Enforcement Agencies; https://copkit.eu/
TRAVELOGUES (FWF) – Perceptions of the Other 1500–1876; http://www.
travelogues-project.info/
Pelagios Projects – a collection of multiple projects under which the Recogito
platform was developed: http://commons.pelagios.org/ (Andrew. W. Mellon
Foundation), https://pelagios.org/
Using pre-defined or manually assigned labels, classification
algorithms support the structuring and filtering of large text
collections into sub-groups. This is also important for building
a domain-specific, cleaned corpus for use in a specific applica-
tion domain. Once such a well-defined text corpus exists, it can
be used to implement supervised machine learning tasks.
These include, for example, Named Entity Recognition (NER) in
which, based on external knowledge, words are assigned to
abstract classes such as names of persons or organizations,
events, machine parts in industry or medical terms in healthcare.
AIT´S ANNOTATION PLATFORM
One of our primary goals is to minimize the effort required for
the manual annotation and labelling of text, because this task
relies on manual work and domain knowledge. One method of
achieving this goal is to provide efficient and user-friendly
labelling tools. Recogito is our annotation platform which uses
existing gazetteers and dictionaries to aid in the process of
producing the high-quality ground-truth training data required
for supervised machine-learning tasks. This task can be
supported by active learning, where the learning algorithm
actively suggests significant terms to the user who is then able
to confirm, reject or refine effectively, and by transfer learning,
which allows us to leverage generic language models which
can then be tailored to domain-specific languages and docu-
ment types with significantly less effort.
AIT AUSTRIAN INSTITUTE OF TECHNOLOGY
Dr. Sven SchlarbTel. +43(0) 50550 4179Giefinggasse 4, 1210 Vienna [email protected] www.ait.ac.at/dss
10
MULTI-MODAL AI FOR SECURITY APPLICATIONSAIT has developed a scalable security platform for implemen-
ting multi-modal AI at a market-ready level, and with features
that address the requirements of end users including service
providers and stakeholders such as law enforcement agencies
(LEAs). The core task of the platform is to analyse and prioritize
large volumes of audio-video content delivered, for example, by
witnesses to a terrorist attack such as the Boston Marathon
bombing incident. Experience shows that such incidents result
in several thousands of hours of video material. This develop-
ment relies on several areas of AIT expertise:
Image Analysis: Videos are searched for visual features.
Designated objects are detected and tracked. A flexible analysis
component plug-in layer allows for the simple inclusion of
additional algorithms such as license plate number extraction
or facial recognition.
Audio Analysis: a special focus of the platform is audio analy-
sis. Non-visual events such as gunshots and explosions are
indexed to facilitate search, using methods based on deep
neural networks to detect relevant acoustic events. Additional-
ly, relevant video sequences can be selected to identify other
sequences with similar audio content. This is useful as an aid in
personal identification, for example. Videos with similar audio
will be recorded at the same place and at the same time,
possibly providing a different angle on the same person. We
also offer audio fingerprinting techniques that allow the system
to synchronize multiple data sources based on audio events.
Scalable Data Processing: Managing large volumes of video
data requires the use of appropriate technologies. We have
experience in various scalable architectures including Hadoop,
Spark, and Airflow. Additionally, neural network models
optimized on GPUs, such as the audio models for event
detection and scene classification noted above, are integrated
with this distributed, scalable architecture, which is otherwise
based on commodity hardware.
11
AIT AUSTRIAN INSTITUTE OF TECHNOLOGY
Dr. Ross KingTel. +43(0) 50550 4271Giefinggasse 4, 1210 Vienna [email protected] www.ait.ac.at/dss
References (excerpt)
VICTORIA – Video analysis for Investigation of Criminal and TerrORIst Activi-
ties; https://www.victoria-project.eu/ (EU H2020)
FLORIDA; https://www.kiras.at/gefoerderte-projekte/detail/d/florida/ (funded
under the Austrian KIRAS security programme of the Federal Ministry of
Transport, Innovation and Technology)
A scalable system architecture uses multiple distributed open
source platforms in order to manage these requirements. It is
based on multimedia data, assumed to be available after a
terrorist or criminal attack. Among other data, this will include
videos from eye witnesses as well as videos from public and
private surveillance cameras. These large volumes of multime-
dia data are uploaded into a distributed file system. During
import, image and audio features relevant for the training of
deep learning models are extracted, and other necessary
pre-processing steps are executed. The neural networks used
for audio analysis and video tracking are trained and executed
on integrated GPU clusters. A query interface, based on
standards such as JSON and GraphQL, accesses this database
and serves as the interface to external systems.
Image Processing
Document ImagesVisual Document content
Audio Processing
Acoustic Scene ClassificationAudio Event Detection
Audio Similarity
MMulti--MModal
AAI
Text Processing
Natural Language ProcessingNamed Entity Recognition
Topic Modelling
Speech Processing
Speech-To-TextText-To-SpeechVoice Biometrics
Sensor Data
Time-Series PredictionPredictive Analytics
SemanticsOntologiesKnowledge Graphs
12
MULTI-MODAL AI FOR DOCUMENT MANAGEMENTDocument management systems need to be able to store,
archive, process, and extract information from digital-born as
well as scanned, image-based documents. Documents and
their relevant parts need to be searchable and to be presented
appropriately with respect to their context and document type.
In recent times, AI in general, and machine learning in particu-
lar, has opened up new possibilities for better supporting users
and providing more efficient tools for managing large collec-
tions of heterogeneous documents.
AIT offers a full range of software and expertise, from docu-
ment acquisition to access. Its portfolio comprises modular and
flexible components for creating domain-specific document
processing workflows that are ready to scale up and work on
very large document collections. Multi-modal AI is applied to
this domain in three primary areas.
First, the quality of OCR (optical character recognition) has
improved significantly in recent years, also with the help of AI.
While digital-born and good quality scans can be reliably
transformed into text, challenges remain for low quality scans
or photographs taken from documents. Here multi-modal AI
can help by using a combination of techniques (OCR and Image
Analysis) to achieve better overall performance. At AIT, we use
these techniques to address challenging tasks such as gaining
structured information from tables contained in document
images or extracting semantically meaningful blocks including
the sender, addressee, and subject line of a scanned letter.
Second, AI is now at the heart of “document understanding”.
Domain-specific language and document structure and layout
models are required to reach a good level of performance in
document classification, clustering, information extraction, and
natural language analytics tasks. This in turn requires the
creation and management of adequate training data and, in
many cases, a scalable storage and compute environment for
pre-processing and for the creation of machine learning
models. AIT can advise your organization on how to manage
this new aspect of data and software maintenance.
Third, AI is indispensable for the access and retrieval of
documents, and AIT can support the introduction of this new
approach in your knowledge management landscape. This
includes, for example, search technology that relies heavily on
natural language processing techniques, such as NGrams,
Gazetteers and dictionaries, and on Named Entity Recognition,
where we leverage well-known open source search and
retrieval frameworks such as SolR and Elastic Search.
References (excerpt)
E-ARK – European Archival Records and Knowledge Preservation; https://
www.eark-project.com/ (EU CIP), https://e-ark4all.eu/
AIT AUSTRIAN INSTITUTE OF TECHNOLOGY
Dr. Sven SchlarbTel. +43(0) 50550 4179Giefinggasse 4, 1210 Vienna [email protected] www.ait.ac.at/dss
13
MULTI-MODAL AI FOR CUSTOMER SERVICE
Customers are the most important asset in any industry.
Customer service is the art of fulfilling your customer’s
expectations by delivering professional, helpful, high quality
service and assistance before, during, and after the customer’s
requirements have been met.
Here at AIT, we are exploring Artificial Intelligence approaches
that combine multiple data types, such as images, audio
streams and text, so-called “multi-modal AI”, so that your
business can provide an enhanced and efficient customer
service.
For example, how can you automatically determine who is the
sender or recipient of a letter when technologies such as
Named Entity Recognition (NER) provide only names? In this
case more context is required. Often the sender uses a diffe-
rent font, a certain style or even a logo in the letterhead. This
layout information is not available in pure text data, but by
combining text analytics with visual analytics, the text is set
within a broader context and the two names can be distinguished.
Multi-modal AI combines information from multiple modalities,
such as images, audio, text, sensors, etc. Traditionally complex
to realize, this field is currently experiencing rapid advances
due to the success of deep neural networks which
facilitate multiple modalities within a single model.
With teams specialized in the individual modalities, we work
together intensively to improve our multi-modal approaches
and redefine the state-of-the-art.
References (excerpt)Text and speech recognition software for international provider of enterprise content management software and financial service provider
AIT AUSTRIAN INSTITUTE OF TECHNOLOGY
DI Alexander SchindlerTel. +43(0) 50550 2902Giefinggasse 4, 1210 Vienna [email protected] www.ait.ac.at/dss
14
AI FOR PREDICTIVE MAINTENANCE IN INDUSTRY 4.0Prospective maintenance refers to a holistic approach to
monitoring the history and condition of industrial equipment in
order to determine when maintenance activities should be
carried out. This approach promises cost savings compared to
routine or time-based preventive maintenance regimens,
because maintenance tasks are performed only when required
and expensive post-breakdown interventions are avoided.
AIT trains and applies machine-learning models to detect
degradations of machines or machine parts, allowing them to
predict future equipment failures. This provides an informed
basis for proactive decisions by maintenance planning and can
thus reduce the costs related to unexpected machine break-
downs. For example, maintenance can be brought forward, or
additional resources can be requested. An additional cost
saving occurs when maintenance is avoided for machines in a
healthy condition.
Creating a forecasting model requires comprehensive analysis
and processing of existing data. In this step, AIT experts gather
information from different data sources that are then standardi-
zed, visualized and statistically analysed. The data quality is also
assessed, and our customers receive an analysis of this data
quality and methods for the systematic examination of the data.
The explorative analysis is used to create and evaluate machi-
ne-learning models. Of particular note are "deep learning"
approaches, which yield excellent results in many cases, and
classical models that allow interpretation. Interpretability is
desirable in many applications because decisions become
human-readable and confidence in the model is strengthened.
Together with the domain expert, the current maintenance
strategy is depicted in realistic benchmarks. The evaluation of
the models is based on real customer data and in reference to
these benchmarks. In order to estimate the effort of an imple-
mentation to production, a deployment concept is created based
on the requirements of the forecast model and the customer’s
local infrastructure. Our expertise and offering:
• Data analysis & visualization operating on top of data streams
• Data quality checking
• Artificial Intelligence models predicting time-to-failure (TTF)
and remaining useful lifetime (RUL)
• Automatic health monitoring (health factor generation)
• Building decision support models that consider algorithmic
predictions
• Performance metrics, test methods, data sets, and referen-
ce documentation to verify and validate deployed methods
References (excerpt)
Machine learning and deep learning tools for international technology manu-
facturers and technology companies
COGNITUS (FFG IKT der Zukunft) – Deep learning technology for predicting
outages of machineries based on sensor data streams
AIT AUSTRIAN INSTITUTE OF TECHNOLOGY
Dr. Bernhard HaslhoferTel. +43(0) 50550 4192Giefinggasse 4, 1210 Vienna [email protected]/dss
15
AI FOR SAFETY & SECURITY IN CYBER-PHYSICAL SYSTEMS Artificial Intelligence, together with safety & security require-
ments, is significantly shaping the systems used in Industry 4.0,
highly automated cyber-physical systems, and the Internet of
Things (IOT). The Center for Digital Safety & Security has
extensive experience in the development and operation of safe
and secure systems. For many years specialists at the Center
have been working in standardization in order to shape tomor-
row’s compulsory standards, as well as developing practical
solutions and advising companies on the introduction of
compliant processes and implementing engineering projects.
The Center’s portfolio includes:
• Training in Artificial Intelligence (AI) for industrial applica-
tions, cybersecurity, standard-compliant work
• Verification technologies suitable for AI methods, and which
themselves apply AI approaches in order to increase
efficiency
• Standard-compliant workflows including tool support, e.g.
using AI-based document evaluation
• Tools for automatic, model-based threat analysis of system
designs (Security by Design) with an automatically updating
threat database (subscription model)
• Real-time monitoring of analogue and digital signals
(runtime monitoring)
• Data analysis and machine learning (including explainable
AI)
• Secure architectures for legacy systems (secure gateways,
etc.)
• Security analyses of network designs, machine code
analysis, source code reviews/audits, penetration testing,
ISO training
• Tools for automated cybersecurity threat analysis in the
automotive field
The special processes, technologies and tools developed at AIT
allow the safety & security of systems to be monitored in
real-time, during operation, and are currently being deployed
globally.AIT AUSTRIAN INSTITUTE OF TECHNOLOGY
Dr. Willibald KrennTel. +43(0) 50550 4109Giefinggasse 4, 1210 Vienna [email protected] www.ait.ac.at/dss
References (excerpt)
Enable-S3 – European Initiative to Enable Validation for Highly Automated
Safe and Secure Systems; https://www.enable-s3.eu/ (H2020, ECSEL Joint
Undertaking)
THREATGET – Threat Analysis and Risk Management Tool; https://www.
threatget.com/, https://cybersecurity.lieberlieber.com
Installations at Tier 1 manufacturers and automotive suppliers
AIT AUSTRIAN INSTITUTE OF TECHNOLOGY
Dr. Markus KommendaBusiness DevelopmentCenter for Digital Safety & SecurityPhone +43 50550 4180 Giefinggasse 4, 1210 Vienna, [email protected]/dss
Ed
ito
ria
l m
an
ag
em
en
t: M
ich
ae
l W
. M
ürli
ng
| Im
ag
es:
iSto
ck
.co
m, A
IT, J
oh
an
ne
s Z
inn
er
| P
lac
e o
f p
ub
lic
ati
on
: V
ien
na
, A
ug
us
t 2
019
AIT AUSTRIAN INSTITUTE OF TECHNOLOGY
Mag. (FH) Michael W. MürlingMarketing and CommunicationsCenter for Digital Safety & SecurityPhone +43 50550 4126 Giefinggasse 4, 1210 Vienna, [email protected]/dss